UNet

Overview

First solution uses U-Net as a model. The pipeline is defined in the pipelines.py:L121 (training) and pipelines.py:L180 (inference). It is end-to-end solution, which means that it takes raw input data and returns Kaggle submit. First solution consist of following steps:

Preparations: perform train-validation split

Pipeline steps:

1. metadata preprocessing (for training purposes)
2. preparing data-loader (PyTorch)
3. unet_network -> define and train U-Net. Network can be parametrized in the: neptune.yaml:L43
4. mask resize -> resize model's predictions to original size
5. thresholder -> binarize masks

prepare Kaggle submit based on pipeline output.

Run default experiment

Run command:

neptune login
neptune send main.py --worker gcp-gpu-medium --config neptune.yaml --environment pytorch-0.2.0-gpu-py3 -- train_evaluate_predict_pipeline --pipeline_name unet